A Feasibility Study on Low Level Techniques for Improving Parsing Accuracy for Spanish Using Maltparser

In the last years dependency parsing has been accomplished by machine learning–based systems showing great accuracy but usually under 90% for Labelled Attachment Score (LAS). Maltparser is one of such systems. Machine learning allows to obtain parsers for every language having an adequate training corpus. Since generally such systems can not be modified the following question arises: Can we beat this 90% LAS by using better training corpora? Some previous work points that high level techniques are not sufficient for building more accurate training corpora. Thus, by analyzing the words that are more frequently incorrectly attached or labelled, we study the feasibility of some low level techniques, based on n–version parsing models, in order to obtain better parsing accuracy.

• Type: Conference
• Year: 2010
• Where: SETN
• Authors: Miguel Ballesteros, Jesús Herrera, Virginia Francisco, Pablo Gervás
• Keywords: Artificial Intelligence, Better Parsing Accuracy, Parsing, SETN 2010, Training Corpora